Skip to main content

​
Pipeline

Every research session runs through a 6-block pipeline: 
You: /autosearch "topic"
 |
 v
[1] Claude recalls what it already knows -> maps 9 knowledge dimensions
 |
[2] Identifies gaps -> generates queries ONLY for what Claude doesn't know
 |
[3] Searches 32+ channels in parallel (10-30 seconds)
 |
[4] LLM evaluates each result for relevance, filters noise
 |
[5] Synthesizes report with two-stage citation lock
 |
[6] Checks quality rubrics -> evolves strategy -> commits improvements
Each block has a time limit (5/5/5/8/3/3 minutes) to prevent hangs.

​
Self-evolution (AVO)

AutoSearch implements the AVO (Agent Variation Operator) pattern:
  • P_t = state/ — lineage: worklog, patterns, evolution history, outcomes
  • K = skills/ — capability set (50+ skills)
  • f = judge.py — fixed scoring function (never modified by evolution)
After each session:
  1. judge.py scores results on 7 dimensions
  2. If the score improved, the strategy changes are committed
  3. If the score regressed, changes are reverted
  4. Reusable patterns are saved to state/patterns.jsonl
The safety mechanism: the evaluator is fixed and cannot be modified by evolution. Only search strategy evolves — not the scoring.

​
Model routing

TaskModelWhy
Classification, scoring, rubricsHaikuFast, cheap, sufficient
Synthesis, reasoning, evolutionSonnetBetter judgment
Claims compressionHaikuBatch processing

​
Directory layout

autosearch/
  .claude-plugin/     Plugin manifest (version, metadata)
  commands/           /autosearch entry point + setup
  agents/             Claude Code agent definitions
  skills/             50+ skills (pipeline, synthesis, evaluation, evolution)
  channels/           32+ channel plugins (SKILL.md + search.py each)
    _engines/         Shared backends (baidu, ddgs)
  lib/                search_runner.py, judge.py (fixed evaluator)
  state/              Append-only learning data
  scripts/            Setup, install, version management
  tests/              pytest test suite (279 tests)
  evidence/           Search results (per-session)
  delivery/           Generated reports
  docs/               This documentation site

​
Immutable files

These files cannot be modified during normal operation:
  • PROTOCOL.md — agent operating protocol
  • lib/judge.py — scoring function
  • Meta-skills: create-skill, observe-user, extract-knowledge, interact-user, discover-environment